Identifying Novel Phenotypes of Renal Recovery and Progression in Sepsis-Associated Acute Kidney Injury: A Dynamic Time Warping Clustering Study in MIMIC-III

Purpose: Sepsis-associated acute kidney injury (SA-AKI) is a grave condition with high morbidity. The current understanding of renal recovery is often limited to binary outcomes, which fail to capture the complex and heterogeneous temporal dynamics of kidney function. This study was undertaken to identify distinct, clinically meaningful phenotypes of renal function trajectories in patients with SA-AKI. Methods: We conducted a retrospective cohort study using the MIMIC-III v1.4 database. Adult patients with a diagnosis of both sepsis (using Sepsis-3 criteria) and AKI (based on KDIGO guidelines) were included. Time-series serum creatinine (SCr) and urine output (UO) data for the first 14 days following AKI diagnosis were extracted. We employed a multivariate Dynamic Time Warping (DTW) distance metric coupled with K-medoids clustering to group patients based on the similarity of their renal function trajectories. The optimal number of clusters was determined using the average silhouette score. The identified phenotypes were then characterized based on baseline patient characteristics, clinical interventions, and outcomes. Results: From an initial pool of 12,542 patients with sepsis, our final cohort comprised 4,067 patients with SA-AKI. We identified four distinct phenotypes: Phenotype 1 'Rapid Reversal' (n=1430, 35.2%), Phenotype 2 'Gradual Recovery' (n=828, 20.4%), Phenotype 3 'Fluctuating Non-Recovery' (n=1293, 31.8%), and Phenotype 4 'Progressive Worsening' (n=516, 12.7%). While 90-day mortality was similar across phenotypes (ranging from 20.7% to 22.5%, p=0.893), baseline severity and clinical outcomes varied significantly. The 'Progressive Worsening' phenotype had the highest baseline SOFA score (13.6 ± 3.8) and the highest rate of RRT initiation (72.1%), in stark contrast to the 'Fluctuating Non-Recovery' phenotype (SOFA 9.0 ± 3.7, RRT 3.9%). The 'Rapid Reversal' and 'Gradual Recovery' groups showed intermediate severity and RRT rates. Conclusion: Our study demonstrates that a data-driven approach using DTW clustering can successfully unravel the significant heterogeneity in renal recovery patterns among SA-AKI patients. The four identified phenotypes, while not differing in 90-day mortality, represent distinct clinical pathways characterized by different levels of illness severity and need for organ support. This nuanced phenotyping could pave the way for more personalized prognostication and the design of targeted therapeutic strategies in the future.